tree e4357bbcc0c0fe7b31b49349c6c37054307b31dd
parent 81ddc344bdb4f93cfe0b498c35a05f0b70d6a9c9
author Marc Kupietz <kupietz@ids-mannheim.de> 1788434622 +0200
committer Marc Kupietz <kupietz@ids-mannheim.de> 1788436152 +0200

Doc prompting tests: cover authorization and multi-VC comparison

The two Readme sections that are hardest to follow were not covered by
any documentation task: authorization, which readers most often get
wrong, and the multi-VC collocation comparison, whose section is new and
whose columns are easy to misread.

Unlike the other tasks, the code of these two cannot reasonably be
executed in a test: authorization needs a browser flow or a token for
restricted data, and a multi-VC collocation analysis runs for minutes.
Only the generated code is inspected, which is what these tests are
about anyway - whether the Readme can be followed.

The expectations therefore aim at the parts that are easy to get wrong
and that the sections explicitly explain: metadataOnly = FALSE for
receiving KWIC snippets, a *named* vc vector, since the names become the
labels of the comparison columns, reducing the result to one row per
collocate, since it holds one per collocate and vc, and taking the
imputed flag into account, since imputed scores describe presence or
absence rather than a measured contrast.

GLM-5.3-Flash solves both from the Readme alone, in the multi-VC case
including filter(label == ..., !imputed) and arranging by delta_logDice.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Change-Id: I25865ef32bb6c3ee894cfb081f93d51bfb28b915